The HRE-NDC utilizes a multi-scale structure, as illustrated in the figure. The user can choose the scale as per their requirements. Unlike other multi-scale networks used for recognition or classification tasks, which extract features at different scales and concatenate them into one, our approach aims to minimize computation at high-resolution scales to enhance efficiency. Therefore, we downsample the ground truth to various scales to generate corresponding pseudo targets. Assuming that each downsampling reduces the grid to half of the current resolution, we represent the number of downsampling times as "j," and the lowest scale in the figure has undergone two downsamplings (j=2). These pseudo targets can be denoted as:

It is important to note that our approach differs from others in that we supervise the network at the scale where the ground truth is located to improve the prediction effect.


原文地址: https://www.cveoy.top/t/topic/p42 著作权归作者所有。请勿转载和采集!

免费AI点我,无需注册和登录